Tree Pruning Puzzle Game

Game setup and pruning laboratory

Build a noisy tree, inspect branches, preview consequences, then prune carefully.

Ready Build 2026-08-11

Fictional dataset

Game mode

Difficulty

Pruning strategy

Dataset size 180

Label noise 12%

Positive-class share 50%

Full-tree depth 6

Complexity penalty alpha 0.018

Maximum target depth 4

Minimum samples per leaf 8

Minimum samples per split 16

Minimum impurity reduction 0.025

Allowed validation loss 2.0%

Validation tolerance 1.0%

Pruning aggressiveness 45%

Training accuracy

—

Original baseline

Validation accuracy

—

Primary target

Test accuracy

—

Generalisation check

Active nodes

—

Complexity

Tree depth

—

Longest route

Score

Pruning quality

Original tree versus pruned tree

Select an internal node on the right. Review its preview before confirming removal.

Split Leaf Pruned Hint

Current objective

Moves: 0

Generate a round to begin.

Your progress will appear here.

Selected branch preview

Select a blue internal node on the pruned tree.

No pruning preview is active.

Coach feedback

Prune branches that add complexity without helping validation performance.

Keyboard: P prune, U undo, H hint, R reset.

Move history

0 badges

No moves yet.

Model evidence and pruning graphs

Plotly charts update after every pruning action.

Before-and-after model comparison

Metric	Original	Current	Change

Precision

—

Recall

—

F1 score

—

Generalisation gap

—

Rating: Waiting for your first pruning decision.

Results and export tools

Generate a round, then simplify the model while protecting validation accuracy.

Local storage keeps scores, badges, settings, and the latest tree state in this browser.

How pruning decisions work

1. Inspect

Check samples, impurity, information gain, depth, and validation impact before removing a subtree.

2. Preview

The preview temporarily replaces a branch with its majority-class prediction and calculates changed metrics.

3. Balance

Strong pruning reduces variance. Excessive pruning raises bias and can create underfitting.